OptiMAS: Automatically Optimize Multi-Agent System

AI summary

Overview

• Research area: Multi-Agent Systems (cs.MA) — automated design and evolution of LLM-based multi-agent architectures.
• Technical level: Advanced. The paper assumes familiarity with ReAct agent loops, tool-calling LLM agents, orchestration graphs, and search-based vs. optimization-based evolutionary methods.
• Scope in one sentence: The paper proposes an optimization-based paradigm and a task-agnostic "agentic optimizer" (OptiMAS) that continuously evolves a multi-agent system's prompts, skills, orchestration topology and toolkits, guided by textual "gradients" derived from execution trajectories, and evaluates it on four agentic benchmarks with three LLM backbones.

What This Paper Is About

Building a multi-agent system (MAS) today is mostly manual work: someone must calibrate prompts, add capabilities, and design the coordination structure through repeated trial and error. Existing automated approaches search over candidate designs in a discrete, branch-and-discard fashion, which has two problems: expanding what can be optimized makes evolution unstable, and insights learned in one search branch are thrown away when that branch is discarded. The paper reframes MAS evolution as a continuous, data-driven optimization process — analogous to training a neural network — and builds OptiMAS, an optimizer agent that reads the evolving system's execution trajectories and accuracy scores, diagnoses failures, and rewrites the MAS step by step.

Key Contributions

1. An optimization-based MAS evolution paradigm. Instead of selecting candidates from a tree of branches, evolution is reformulated as continuous refinement guided by textual gradients, with strict train/validation/test partitioning to control overfitting.
2. A unified ReAct-based infrastructure with a proactive assign-deliver protocol. Every agent is a recursive ReAct reasoner; delegation to sub-agents is modeled as a tool-invocation action, so agents are activated on demand rather than by a hardcoded workflow. The optimization scope covers the agent population, the orchestration graph (constrained to a DAG), and each agent's LLM backbone, prompts, skill set and toolkit — with tools drawn from a predefined macro-library rather than generated as code.
3. OptiMAS, a task-agnostic agentic optimizer with a dual-track memory mechanism. Plan-oriented short-term memory enforces intra-step consistency (four instruction statuses: waiting, executing, completed, dropped; four operations: add, modify, review, delete); hypothesis-driven long-term memory maintains a cross-step ledger of falsifiable hypotheses with a Propose → Enact → Validate/Refute/Suspend lifecycle.
4. Comprehensive evaluation on four heterogeneous agentic benchmarks with three backbones of different scale and accessibility, compared against both hand-crafted domain-specialized MAS and existing evolutionary MAS methods.

Main Findings

• OptiMAS improves on its own starting point everywhere tested. A single task-agnostic OptiMAS configuration — no task-specific adaptation — is applied across all four benchmarks and three backbones, with gains ranging from +2.8% (BrowseComp/GPT5) to +45.4% (WorkBench/GPT5) over the single-agent initialization ℰ₀.
• Per-benchmark accuracy (Table 1). WorkBench: 64.0 (GPT5), 84.5 (Qwen3.6), 92.9 (Gemini3). GAIA: 56.9, 81.2, 87.1. BrowseComp: 11.1, 43.8, 58.3. SWE-Bench: 41.7, 68.9, 77.2. Corresponding ℰ₀ baselines are 18.6/65.6/71.7 (WorkBench), 49.5/73.1/80.7 (GAIA), 8.3/33.9/52.2 (BrowseComp) and 29.4/61.1/72.8 (SWE-Bench).
• Cross-domain generalization. On GAIA and BrowseComp, OptiMAS surpasses the domain-specialized hand-crafted Tongyi-DR with Qwen3.6 (81.2%/43.8% vs. 80.7%/38.3%) and Gemini3 (87.1%/58.3% vs. 75.8%/48.3%). On SWE-Bench it outperforms SWE-Agent across all three backbones. On GAIA with GPT5, however, OptiMAS reports 56.9% against Tongyi-DR's 62.9%.
• Large margin over prior automated methods. The paper reports that ADAS's evolved systems fall below OptiMAS by up to 58.9% on SWE-Bench/Gemini3.
• OptiMAS is the best result in most, not all, benchmark–backbone configurations. GAIA/GPT5 is the explicit exception noted above.
• Gains are larger on weaker backbones. The evolution gain is reported as inversely correlated with backbone capacity, e.g. +45.4%/+12.3% on WorkBench/SWE-Bench under GPT5 versus +15.5%/+4.4% under Gemini3 as stated in the text (Table 1 lists the WorkBench/Gemini3 gain as +21.2%).
• Evolution dynamics are smooth and near-monotonic. Over 10 optimization epochs on WorkBench, training accuracy gains exceed +43% (GPT5), +36% (Qwen3.6) and +21% (Gemini3) absolute, with test accuracy closely tracking and a narrow train–test gap throughout.
• Ablations identify the memory mechanism as most critical. Removing hypothesis memory drops SWE-Bench below ℰ₀ on both backbones (−7.2% GPT5, −7.8% Qwen3.6). Removing adaptive sampling also falls below ℰ₀ on SWE-Bench (−5.0%/−0.5%). Cutting batch size from 10 to 4 costs GPT5 −5.0% on SWE-Bench. Removing the validation metric has a moderate overall effect but introduces slight overfitting on SWE-Bench/GPT5 (−0.5%).
• The optimizer's own backbone matters most on trajectory-heavy tasks. Substituting Qwen-3.6-35B-A3B as the optimizer yields positive gains over ℰ₀ in all eight tested configurations, but only marginal gains on GAIA and BrowseComp (+1.6%/+1.1% for Qwen-MAS) where the Gemini optimizer achieves +8.1%/+9.9% — attributed to a context-length bottleneck.
• Token cost. OptiMAS processes 1.4M to 3.2M tokens per optimization step across the four benchmarks (roughly 750K to 2.4M words), with input constituting over 98% of consumption. The evolved system ℰ* is then deployed at standard inference cost.

Methodology in Plain English

The paper borrows the shape of deep learning training and applies it to agent architecture design.

1. Make everything configurable. Every agent is a standard reason-act loop (ReAct). Agents ask each other for help by calling a "assign" tool; the helper answers and its answer comes back as an observation. Nothing about the workflow is hardcoded. Because everything is configuration, the thing being optimized is a set: the agents, the directed acyclic graph connecting them, and for each agent its backbone, prompt, skills and tools.
2. Split the data three ways. Queries are divided into a train set, a small validation set, and a held-out test set. The evolving system runs on train queries each step and its full workspace — every trajectory, code snippet, retrieved document and working note — is handed to the optimizer. It also runs on the validation set, but the optimizer only ever sees the scalar accuracy number, never the validation workspace, so that score acts as a generalization check rather than something to fit against. The test set is used only for final reporting.
3. Convert failures into "textual gradients." The optimizer inspects the workspace, diagnoses root causes of failure, spots recurring success patterns, and directly rewrites the MAS configuration. Gradient analysis and the parameter update happen in one end-to-end pass rather than as separated stages, which the authors argue keeps the reasoning coherent.
4. Sample smarter. Adaptive sampling with priority weights revisits queries the system previously failed while still exploring the rest of the training set, so the optimizer keeps getting signal from its known weaknesses.
5. Remember across steps. Short-term memory holds a plan with explicit statuses so the optimizer does not drift or forget mid-step, and instructions can be "dropped" only with written justification — content can never be deleted, preserving an audit trail. Long-term memory stores each proposed change as a falsifiable hypothesis with expected outcomes; hypotheses move through Enact, Validate, Refute or Suspend states based on trajectory-level evidence, and every edit to the MAS must cite a hypothesis ID.
6. Compare fairly. All evolution baselines use the same Gemini-3-Flash meta-agent, the same data subsets and splits, and each benchmark's official metric protocol (sandboxed execution for SWE-Bench-Verified and WorkBench, an LLM-as-a-Judge pipeline for BrowseComp and GAIA-text), with test results averaged over three independent trials.

Why This Matters

The paper argues that hand-crafted MAS still dominate industrial practice because automated alternatives either destabilize when given a wide design space or waste insights by discarding search branches. OptiMAS is a step toward automated synthesis of deployable multi-agent systems from a single task-agnostic optimizer, and its results suggest the evolved orchestration can partly compensate for a weaker underlying LLM.

Real-world applications:

• Workplace task automation — WorkBench covers analytics, calendar, CRM, email, project management and multi-domain tasks requiring multi-step planning and tool manipulation.
• Deep research and web investigation — BrowseComp and GAIA target long-horizon web navigation, multi-page synthesis, file parsing and cross-tool orchestration.
• Software engineering — SWE-Bench-Verified tasks agents with resolving real GitHub issues through multi-file code comprehension and repository modification.
• Enterprise MAS deployment — the paper frames token-scale workspace analysis (1.4M–3.2M tokens per step) as meeting industrial MAS evolution requirements, with the evolved system deployed at ordinary inference cost.

Industry relevance: the work is co-authored with Huawei's Foundation Model Department and reports results spanning a closed-source model, an open-source model and a commercial frontier model, which speaks directly to practical deployment where the underlying LLM may not be fixed or freely chosen.

Future Directions

• Lowering backbone requirements. The infrastructure's demands on instruction-following and tool-calling mean substantially weaker models may struggle; the authors propose simplified execution protocols or progressive capability scaffolding for lighter-weight or edge-friendly models.
• Broadening the optimization search. Training curricula and scheduling, dataset curation policies, multi-agent diversity and composition, and treating prompt templates and tool creation as first-class optimization targets all remain unexplored.
• Closing the optimizer context bottleneck. The gap between the Qwen and Gemini optimizers on trajectory-heavy benchmarks points to optimizer context capacity as a scaling factor for evolution quality on long-horizon tasks.
• Scaling evaluation. BrowseComp and SWE-Bench-Verified were evaluated on fixed 100-instance subsets (GAIA-text and WorkBench used full sets), leaving the behavior of OptiMAS on the complete benchmark suites an open question.

Target Audience

Researchers and engineers working on LLM agent architectures, automated prompt/workflow optimization, and evolutionary program search will get the most from this paper, particularly those interested in treating agent design as an optimization problem rather than a search problem. Practitioners who need to stand up multi-agent systems for research, software engineering or web-information tasks will find the evaluation setup and the ablation results on memory and sampling directly actionable. Readers new to agent systems will need prior familiarity with ReAct-style agents and multi-agent orchestration to follow the methodology.

Authors’ abstract

Automated evolution of Multi-Agent Systems (MAS) holds significant potential for reducing the manual effort required to design and optimize LLM-based agent architectures. However, extant search-based paradigms face a fundamental trade-off, where an expanded optimization scope exacerbates evolutionary instability, while discrete branch-and-discard search isolates insights across lineages. To address these limitations, we propose a continuous, data-driven optimization paradigm built upon a unified ReAct-based infrastructure that reconciles a broad optimization scope with operational stability. Under this paradigm, we present OptiMAS, a task-agnostic agentic optimizer that leverages textual interaction trajectories and task feedback as loss signals for end-to-end MAS evolution. Equipped with a novel dual-track memory mechanism, OptiMAS sustains performance improvement over extended optimization horizons. Evaluation on four heterogeneous agentic benchmarks with three varying scale and accessibility LLM backbones, demonstrates that OptiMAS consistently achieves competitive or superior accuracy relative to both domain-specialized hand-crafted systems and existing evolutionary methods. Our work establishes a practical milestone toward robust, automated MAS evolution.